The liquid is part of the colour: what AI learns about fluorescent dyes

One molecule, two liquids, two different emission peaks. In a 2019 experiment, a fluorescent probe called BTD-4APTEG emitted most strongly at 517 nanometres in toluene and 543 nanometres in water, at the same measurement concentration. Its molecular structure had not changed. Its surroundings had.[1]

The probe belongs to the benzothiadiazole family, usually shortened to BTD. Carvalho and colleagues developed it to illuminate cell membranes. Their calculations traced a redistribution of electronic charge when light excited the molecule; the solvent helps determine the energies involved. The colour a chemist wants to predict therefore belongs to a molecule in an environment, not to a structural drawing alone.[1]

That distinction gives a new Scientific Reports study, published on 17 October as an accepted manuscript, a practical purpose. Duc Toan Truong and colleagues ask how well machine learning can predict the wavelengths BTD dyes absorb and emit when their structures and solvents vary.[2]

A familiar dye in a different liquid

The earlier Veríssimo team’s public modelling file contains 701 rows, but only 174 distinct molecular structure strings and 26 solvent labels. We checked those counts directly. A row is a molecule–solvent observation: several rows can describe the same dye. BTD-4APTEG, for example, appears with six solvents.[3]

This creates two different prediction jobs. A chemist may already have a promising dye and want to choose a liquid for it. Or the chemist may want to make a new molecular structure. A test that withholds one measurement of a familiar molecule answers the first question more directly than the second.[3]

In our illustrative 80:20 shuffle of the public file, 129 of 141 held-out rows—91.5%—had a structure also present in the remaining rows. This is a reproducible data check, with no model trained. It shows how easily a large-looking test set can remain chemically familiar.[3]

Truong’s team used molecule-grouped hold-outs, keeping all measurements of a held-out structure out of training. The supplementary results put average absolute errors for the leading tree models at 21.0 nanometres for absorption and 30.3 nanometres for emission across five such splits. Each split withheld 35 structures.[4]

Give the model the chemistry it needs

The researchers also improved the molecular description by counting repeated chemical substructures. A yes-or-no fingerprint can record that an electron-donating group exists while losing how often it occurs. The count representation retains that information. Their model analyses found solvent polarity much more influential for emission than for absorption—a pattern consistent with the way the excited molecule and its surroundings relax before light is emitted.[2]

An imaging dye also has to reach the desired part of a cell and behave usefully under illumination. The membrane experiments in the 2019 study addressed those properties directly. Wavelength prediction answers one part of the design problem; the laboratory has to supply the rest.[1]

A useful next experiment may simply be a change of solvent. The molecule on the bench can stay the same.

Report an error or suggest a correction